See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/256202854

Simulation of Hyperbolic Stress-Strain Parameters of Soils Using Artificial

Neural Networks

Conference Paper · February 2013

CITATION READS

1 425

3 authors:

Khalid Rassim Aljanabi Adel H. Majeed

University of Anbar Al-Mustansiriya University
18 PUBLICATIONS   8 CITATIONS    1 PUBLICATION   1 CITATION   

SEE PROFILE SEE PROFILE

Ameer A Jebur
Liverpool John Moores University
17 PUBLICATIONS   17 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Prediction of Ultimate Bearing Capacity of Shallow Foundations on Cohesionless Soils Using Back Propagation Neural Networks (BPNN) View project

Modeling of Polymer Modified-Concrete Strength with Artificial Neural Networks View project

All content following this page was uploaded by Khalid Rassim Aljanabi on 20 May 2014.

The user has requested enhancement of the downloaded file.

International Conference on Geotechnical Engineering. 2013

Simulation of Hyperbolic Stress-Strain Parameters of Soils Using

Artificial Neural Networks

Dr. Adel H. Majeed

Assistant Professor, Civil Engineering Dept., College of Eng., AL-Mustansiriya University-Iraq

Dr. Khalid R. Mahmood

Assistant Professor, Civil Engineering Dept., College of Eng., University of Anbar-Iraq

Ameer A. Jepur
M.Sc., Civil Engineering Dept., College of Eng., AL-Mustansiriya University-Iraq

ABSTRACT: Artificial Neural networks (ANNs) are one of the widely used modeling techniques
during the last decade or so, which can approximate a non-linear relationship between input and
output data sets. The use of artificial neural networks (ANNs) has increased in many areas of
engineering. In particular, ANNs have been applied to many geotechnical engineering problems
and have demonstrated some degree of success. In the present investigation, an attempt is made to
predict hyperbolic stress-strain relationship parameters using multilayer perceptrons that are trained
with back-propagation algorithm for estimating modules number (k), modules exponent (n) and the
failure ratio (Rf). A database comprises a total of (83) case recorded are used to develop the ANN
models. In addition, the paper discusses a number of issues in relation to ANN construction such as
the effect of ANN geometry and internal parameters on the performance of ANN models. Practical
equations for prediction of hyperbolic stress-strain relationship parameters are developed. It was
found that ANNs have the ability to predict the hyperbolic stress-strain relationship parameters
with a good degree of accuracy.

1 INTRODUCTION

The mechanical behavior of soils may be modeled at various degree of accuracy. A large number
of constitutive models have been proposed by several researchers to describe various aspect of soil
behavior in detail. However, the more sophisticated a soil model is, the more parameters have to be
selected on the basis of soil investigation data.
The hyperbolic stress-strain relationships proposed by Kondner (1963) and developed by Duncan
and Chang (1970) in an attempt to provide a single framework encompassing the most important
characteristics of soil stress-strain behavior, namely, nonlinearity, stress dependency, and
inelasticity using the data available from conventional laboratory tests such as UU and CD triaxial
compression test.

For undrained condition, there are three parameters involved in the hyperbolic stress-strain
relationship namely modulus number (K), modulus exponent (n), and failure ratio (Rf) in addition
to the shear strength parameters, the cohesion intercept (c) and the angle of internal friction ( ). In
general, laboratory tests needed to determine constitutive model parameters are expensive. The
alternative solution is to predict these parameters using an empirical modeling technique based on
traditional statistical regression or other empirical technique which need more simple soil
properties such as Atterberg limits, water content,…..etc.
Over the last few years, the use of (ANNs) has increased in many areas of engineering. In
particular, ANNs have been applied to many geotechnical engineering problems and have
demonstrated some of success [Shahin et.al. 2003, 2008a; 2008b; Shahin 2003].

In this paper an attempt has been carried to explore the use of Artificial Neural Network (ANN)
models for simulation the “hyperbolic stress-strain relationship parameters” of soil namely,
modulus number (K), modulus exponent (n), and failure ratio (Rf), under unconsolidated undrained
condition without volume change.

A commercial software, NEUFRAME v.4 Neusciences. (2000) is used to developed three ANN
models. Four parameters are considered to have the most significant impact on the magnitude of
hyperbolic stress-strain relationship parameters and are being used as an input to the models. These
include the plasticity index, dry unit weight, water content, and confining stress. The outputs of the
model are modulus number (K), modulus exponent (n), and failure ratio (Rf). Multi-layer
perceptron trainings using back propagation algorithm are used in this work. A number of issues in
relation to ANN construction such as the effect of ANN geometry and internal parameters on the
performance of ANN models are investigated. Information on the relative importance of the factors
affecting the hyperbolic stress-strain relationship parameters are presented and practical equations
for prediction of these parameters are developed. It was found that ANNs have the ability to predict
the hyperbolic stress-strain relationship parameters with a good degree of accuracy. The ANN
models developed to study the impact of the internal network parameters on model performance
indicate that ANN performance is relatively insensitive to the number of hidden layer nodes,
momentum term, learning rate, and transfer functions.

2 BRIEF OVERVIEW OF HYPERBOLIC STRESS-STRAIN PARAMETERS

These relationships were developed by Duncan and Chang (1970) for using in incremental finite
element analysis. In each increment of such an analysis, the stress-strain behavior of the soil is
treated as being linear, and the relationship between stress and strain is assumed to be governed by
generalized Hooke’s law of elastic deformation, which may be expressed as follow (Atkinson and
Bransby1978):
1
1 1 2 3
E
1
2 2 1 3 (1)
E

1
3 31 1 2
E
For the special case of axial symmetry, where 2 = 3 and 2 = 3, Eqs(1) become:

1
1 1 2 3
E
(2)
1
3 (1 ) 3 1
E
Where 1, 2, and 3 are the increment in principal stress. 1, 2, and 3 are the corresponding
increment in principal strain. E and are the Young’s modulus and Poisson’s ratio respectively.
For special condition of no volume change, will be constant equal to 0.5.
The nonlinear stress-strain curves for a number of soils could be approximated reasonably accurate
by hyperbolas like the following one [Kondner (1963) and Kondner and Zelasko (1963)] :

( 1 3 ) (3)
a b
Where:
1
b ( 4)
( 1 )
3 ult

Equation (3) may be rewritten in the following form:

a b (5)
( 1 3)

Where a and b are respectively, the intercept and the slope of the best fit resulting straight line,
Equ.(5). The asymptotic stress value, 1 3)ult., may be related to the compressive strength,
1 3)f, by means of a factor Rf as follows:

( 1 3 f ) Rf ( 1 )
3 ult ( 6)

The value of Rf has been found to be between 0.5 and 0.9 [Wong and Duncan (1974)].The stress-
dependency is taken into account by using an empirical equation proposed by Janbu (1963) as
follows:
n
3
Ei K Pa ( 7)
Pa
Both K and n are dimensionless numbers and are the same for any system of units, the units of Ei
are the same as the unit of Pa, atmospheric pressure.
The variation of 1 3)f with 3 is represented by the familiar Mohr-Coulomb strength relationship,
which can be expressed as follows[Wong and Duncan (1974)]:

2 c cos 2 3 sin
( 1 )
3 f (8)
1 sin
The instantaneous slop of the stress-strain curve is the tangent modulus Et. By differentiating
equation (3) with respect to and substituting the expressions of equations 4, 5, 7, 8, and 9 into the
resulting expression for Et, the following equation can be derived:
2 n
Rf (1 sin ) ( 1 3 ) 3
Ei 1 K Pa * (9)
2 cos 2 3 sin Pa

This equation can be used to calculate the appropriate value of tangent modulus for any stress
conditions, 3 and 1- 3), if the values of the parameters K, n, c, , and Rf are known. The
discussion of the unloading-reloading condition and the volume change condition are beyond the
scope of this paper and can be found in [Wong and Duncan (1974)].

3 BRIEF OVERVIEW OF ARTIFICIAL NEURAL NETWORKS (ANNS)

An artificial neural network is a massively parallel-distributed information-processing system that

has certain performance characteristics resembling biological neural networks of human brain to
simulate the manner in which the brain interprets information as determined by the current
knowledge. Artificial neural networks behave in much case the same manner as biological neural
networks. Many authors have described the structure and operation of ANNs [Zurada (1992),
Caudill (1988) and Garson (1991)]. The description of back-propagation multi-layer perceptrons
(MLPs) is beyond the scope of this paper and can be found in many publications (e.g. Wasserman
(1989) and Fausett (1994)). The typical structure of a MLP model are usually arranged in layers: an
input layer, an output layer and one or more hidden layers, each layer consists of a number of
neuron or nodes, in which each neuron in a specific layer is joined to other neurons via weighted
connections. Each neuron consists of a linear combination of weighted inputs plus bias or threshold
value, which is passed though a non-linear activation function (f(.)) (e.g. a sigmoidal or tanh
functions) to produce the neuron’s output (yi). This process is summarized in Eqs. (10) and (11)
and illustrated in Fig. 1:

IJ wij xi j Summtion (10)

yj f (I j ) Transfer (11)

Where:
Ij = the activation level of node j.
wij = the connection weight between nodes i and j.
xi = the input from node i, i = 0, 1, ……,n.
j = the bias or threshold for node j.
yj = the output of node j, and
f(Ij) = the transform (activation) function.

Fig. (1) Typical structure and operation of ANNs (after Shahin, 2003)

4 MODEL ARCHITECTURE, OPTIMIZATION AND STOPPING CRITERIA

In this study, Artificial Neural Network (ANN) models are developed using the NEUFRAME
computer software [Neusciences. (2000)]. Three models are built to predict the stress-strain
parameters, modules number (k), modules exponent (n) and failure ratio (Rf) respectively. For each
model, four variables are presented to the MLP models as model inputs. These include the
following

Plasticity Index, (PI) %.

Dry unit Weight ( d) kN/m3
Water content ( o) %
Confining Stress (CS) kN/m2

Data processing is very important in using neural nets successfully. It determines what information
is presented to create the model during the training phase. It can be in the form of data scaling,
normalization and transformation. Transforming the input data into some known forms (e.g. log.,
exponential, etc.) may be helpful to improve ANN performance. Thus, the modules number is
transformed to (Logk). The outputs of the model are logarithm of modulus number, modules
exponent and failure ratio. (Logk, n and Rf) respectively.

As cross-validation [Stone (1974)] is used as the stopping criterion for MLP models, the data are
divided into three sets, namely: training, testing and validation. When dividing the data into their
subsets, it is essential to check that the data used for training, testing and validation represent the
same population, as recommended by Masters (1993). The statistics of the data used for three
models are presented in Table 1. In total, 80% of the data are used for training and 20% are used
for validation. The training data are further divided into 70% for the training set and 30% for the
testing set. Once the available data have been divided into their subsets, the input and output
variables are pre-processed by scaling them between 0.0 and 1.0 to eliminate their dimension and
to ensure that all variables receive equal attention during training [Masters (1993)]. A simple linear
mapping of the variables’ practical extreme to the neural networks’ is adopted for scaling, as it is
the most common method for this purpose [Masters (1993)]. As part of this method, for each
variable, x, with minimum and maximum values of xmin and xmax, respectively, the scaled value, xn,
is calculated as follows:
x xmin
xn (12)
xmax xmin
Table-1 ANN input and output statistics for the three models (Logk, n and Rf)
Input Variables Output
Plasticity Dry Unit Water Confining
Model Statistical Log k, n,
Index(PI), Weight d, Content Stress,
Name Parameters Rf
% kN/m3 o% kN/m2
Maximum 45.000 20.17 31.1 1452.38 3.949
minimum 1.000 10.06 8.3 107.584 1.000
Logk mean 18.807 16.509 17.236 394.608 2.284
Std.dev 8.843 1.686 5.099 224.344 0.649
Range 44.000 10.11 22.8 1344.796 2.949
Maximum 45.000 20.17 31.1 1452.38 1.43
minimum 1.000 10.06 8.3 107.584 -1.1
n mean 18.807 16.509 17.236 394.608 0.190
Std.dev 8.843 1.686 5.099 224.344 0.447
Range 44.000 10.11 22.8 1344.796 2.53
maximum 45.000 20.17 31.1 1452.38 1.000
minimum 1.000 10.06 8.3 107.584 0.52
Rf mean 18.807 16.509 17.236 394.608 0.847
Std.dev 8.843 1.686 5.099 224.344 0.104
Range 44.000 10.11 22.8 1344.796 0.48
To examine how representative the training, testing and validation sets are with respect to each
other, t-test and F-test are carried out on different combinations of data sets until they are
statistically consistent.

One of the most important and difficult tasks in the development of ANN models is determining
the model architecture (i.e. the number and connectivity of the hidden layer nodes). A network with
one hidden layer can approximate any continuo function, provided that sufficient connection
weights are used, [Shahin (2003)].Consequently, one hidden layer is used in this research. The
general strategy adopted for finding the optimal network architecture and internal parameters that
control the training process is as follows: a number of trials is carried out using the default
parameters of the software used with one hidden layer and 1,2,3,…….,9 hidden layer nodes. It
should be noted that 9 is the upper limit for the number of hidden layer nodes needed to map any
continuous function for a network with 4 inputs, [Caudill (1988)] and consequently, is used in this
work.

The network that performs best with respect to the testing set is retrained with different
combinations of momentum terms, learning rates and transfer functions in an attempt to improve
model performance, since the back-propagation algorithm uses a first-order gradient descent
technique o adjust the connection weights, it may get trapped in a local minimum if the initial
starting point in weight space is unfavorable. Using the default parameters of the software, a
number of networks with different numbers of hidden layer nodes are developed and results are
shown in Fig.2, using the performance measures coefficient of correlation (r), Root of Mean Square
Error and Mean Absolute error (MAE), The coefficient of correlation is a measure that is used to
determine the relative correlation and the goodness-of-fit between the predicted and observed, The
RMSE is the most popular measure of error and has the advantage that large errors receive much
greater attention than small errors and MAE eliminates the emphasis given to large error.

Figure (2) shows that, the lowest RMSE value is combined with a hidden layer network with (7, 6,
and 1) nodes for model named (logk), (n) and (Rf) respectively. However, it is believed that, the
network with (1), hidden layer nodes is considered optimal, as the prediction error is not far from
the network with (7) hidden layer nodes for the (logk) model and (6) hidden layer nodes for the (n),
the error difference being only (0.04 and 0.07) respectively, coupled with smaller number of
connection weights.

4
3.5 Logk
RMSEfor testing set

3 n
2.5 Rf

2
1.5
1
0.5
0
0 1 2 3 4 5 6 7 8 9 10
No. of hidden nodes

Figure 2. Performance of the ANN model with different hidden layers

nodes of testing set (Learning rate = 0.2, Momentum term = 0.8)
for (Logk), (n) and (Rf) model

The effect of the internal parameters controlling the back-propagation (i.e. momentum term and
learning rate) on model performance is investigated for the three models with one hidden layer
nodes as shown in figures (3 and 4). It can be seen that the effect of momentum for the three
models are relatively insensitive for momentum term particularly from 0.4 to-1.0, and the effect of
learning rate is insensitive for the three models for values from 0.4-1.0, thus it is decide to use a
network with 1 hidden layer, learning rate = 0.2 and momentum term = 0.8.

3.5

RMSEfor testing set

3 Logk
n
2.5
Rf
2

1.5

0.5

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Momentum term

Fig. (3) Effect of various momentum terms on ANN performance for testing
(Hidden nodes = 1 and learning rate = 0.2), for (Logk), (n) and (Rf) model.

3.5

3 Logk
SEfortestingset

n
2.5
Rf
2

1.5
RM

0.5

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Learning rate

Fig. (4) Effect of various learning rates on ANN performance for testing
(Hidden nodes = 1 and Momentum term = 0.8), for (Logk), (n) and (Rf) model.

This network is trained with different combinations of activation functions and it is found that the
performance of ANN models is relatively insensitive to transfer functions although a slightly better
performance is obtained when the hyperbolic tangent (tanh) transfer function is used for the hidden
layer and the sigmoid transfer function is used for the output layer as shown in Table 2.

Table (2) performance for testing data set (hidden nodes=1 learning rate =0.2 and
momentum term=0.8) under different activation functions
Transfer RMSE for testing data set
Transfer function
function on hidden
on output layer
layer Logk model n model Rf model

Tanh Sigmoid 0.39457 0.31687 0.07064

Sigmoid Sigmoid 0.67461 0.20244 0.08308
Tanh Tanh 0.76060 0.26921 0.09055
Sigmoid Tanh 0.55318 0.41376 0.07521
5 ANN MODELS EQUATION

The small number of connection weights obtained for the optimal model enables the network to be
translated into relatively simple formula:

ANN Model Equation for (Logk):

2.949
Logk 1 (13)
1 e( 0.411 tanh x )

Where:
3
x 10 [(55 PI 310 d 278 n 1.3 3 ) 1708] (14)

ANN Model Equation for (n):

2.53
n ( 0.152 0.623 tanh x )
1.1 (15)
1 e

Where:
3
x 10 [(3.5 PI 80 d 17.7 n 0.2 3 ) 297.7] (16)

ANN Model Equation for (Rf):

0.48
Rf ( 0.753 3.45 tanh x )
0.52 (17)
1 e

Where:
3
x 10 [(3.5 PI 80 d 17.7 n 0.2 3 ) 297.7] (18)

6 VALIDITY OF THE ANN MODELS EQUATION

To assess the validity of the derived equations for the logarithm of modulus number, modules
exponent and failure ratio (Logk, n and Rf) models, the equations are used to predict these values of
all, training, and validation data sets used. Table (3) shows the values of R-squared between the
predicted (observed) and measured values of the (logk, n and Rf), for the three data sets.

Table (3) R-squared values between predicted and measured values of the (logk, n and
Rf), for the three data sets
Data set LogK n Rf
All data 0.73 0.7 0.7
 Training set 0.6 0.7 0.51
Validation set 0.78 0.78 0.66

7 SIMULATION OF STRESS-STRAIN RELATIONSHIP USING ANN EQUATIONS.

Figures (5), and (6), show the comparison between measured stress-strain relationship [Boscardin
(1990)] and predicted stress-strain relationship,based on hyperbolic parameters under different
factros (plasticity index, dry unit weight, water content and confining stress), as shown below.
ANNs, models equations developed to find hyperbolic parameters, logarithm of modules number
(logk), modules exponent (n), and failure ratio (Rf).The figures show a good agreement with the
measured stress-strain relationship.

PI=4 % PI=4 %
3
3 d)=16(kN/m )
d)=18(kN/m )
=12.1 % o =12.1 %
o
CS=104 (kN/m2)
CS=310 (kN/m2)

Fig. (5) Comparison between measured and predicted stress-strain relationship for (ML), soil

PI=15% PI=15 %
3
d)=14(kN/m )
3
d)=15.6(kN/m )
o =21 % o =21 %
CS=310 (kN/m2) CS=104 (kN/m2)

Fig. (6) Comparison between measured and predicted stress-strain relationship for (CL), soil

8 CONCLUSIONS

ANNs have the ability to predict the logarithm of modulus number, modules exponent and
failure ratio (logk, n and Rf), with a good degree of accuracy within the range of data used
for developing ANN models. The models indicate that ANN performance is relatively
insensitive to the number of hidden layer nodes, momentum terms, learning rate and
transfer functions.

9 REFERENCES

Atkinson, J.H. and Bransby, P.L. (1978),” The Mechanics of Soils, An Introduction to
Critical State Soil Mechanics”, McGraw-Hill Book Company (UK) Limited, UK.
 Boscardin, B.D. (1990) “Hyperbolic parameters for compacted clay” ASCE, Vol. 116,
No.1, pp. 88-104.
Caudill M.(1988) Neural networks primer, Part III. , AI Expert, 3(6), 53-59.
Duncan, J.M. and Chang, C.Y. (1970),” Nonlinear Analysis of Stress and Strain in
Soils”, Journal of Soil Mechanics and Foundations Division, Proceedings, ASCE, Vol.
96, No.SM5, pp 1629-1653.
Fausett Laurene V. (1994) “Fundamentals of neural networks: architectures,
algorithms, and application”. Englewood Cliffs, NJ: Prentice-Hall.
Garson, G. D. (1991). “Interpreting neural-network connection weights.” AI Expert,
6(7), 47-51.
Janbu, N. (1963),” Soil Compressibility as Determined by Oedometer and Triaxial
Tests”, European Conference on Soil Mechanics and Foundation Engineering,
Wiesbaden, Germany, Vol.1, pp 19-25.
Kondner, R.L. and Zelasko, J.S. (1963),” A Hyperbolic Stress-Strain Formulation of
Sands”, Proceedings of the 2nd Pan American Conference on Soil Mechanics and
Foundation Engineering, Brazil, Vol 1, pp 289-324.
Kondner,R. L. (1963),” Hyperbolic Stress-Strain Response: Cohesive Soils”, Journal
of Soil Mechanics and Foundations Division, ASCE, Vol.89, No. SM1. Proc. Paper
3429, PP 115-143.
Masters, T. (1993). Practical neural network recipes in C++, Academic Press, San
Diego, California. Diego, California.
Neusciences. (2000). Neuframe Version 4.0, Neusciences Corp., Southampton,
Hampshire.
Shahin M.A, Jaska M.B. and Maier H.R. (2003) “Application of Artificial Neural
Networks in Foundation Engineering”, Australian Geomechanics.
Shahin M.A, Jaska M.B. and Maier H.R. (2008a) “Recent Advances and Future
Challenges for Artificial Neural Systems in Geotechnical Engineering Applications”,
Department of Civil and Environmental Eng., University of Adelaide.
Shahin M.A, Jaska M.B. and Maier H.R. (2008b), “State of the Art of Artificial Neural
Networks in Geotechnical Engineering” Department. of Civil Engineering, Curtin
University of Technology, Perth, WA 6845, Australia.
Shahin, M.A., (2003), Use of Artificial Neural Networks for Predicting Settlement of
Shallow Foundations on Cohesionless Soils, Ph.D. Thesis, Department of Civil and
Environmental Eng., University of Adelaide.
Stone, M. (1974). “Cross-validatory choice and assessment of statistical predictions.”
J.Royal Statistical Society, B 36, 111-147.
Wasserman, P. D. (1989) Neural Computing: theory and practice.
Wong, K.S. and Duncan, J.M. (1974), “Hyperbolic Stress-Strain Parameters for
Nonlinear Finite Element Analyses of Stress and Movements in Soil Masses”, Report
No. TE-74-3, College of Engineering, University of California, Berkeley, California.
Zurada J. M. (1992) Introduction to artificial neural systems, West Publishing
Company, St. Paul.

View publication stats